I sent this email to Anthropic support on 15 October 2026, and I decided to share it as an open letter, because I can't be alone in this. I knew the change to Claude wasn't only hitting me, and some of the findings on GitHub back that up. Their support bot has passed it on to a human, so let's see if there's a response, and what it'll be.
So is this the next case of a frontier AI business, or whatever you want to call it, turning its back on the loyal users who helped train it, build it and get its models to where they are today? I can't help wondering whether this is another brutal change in direction, an effort to force its highest-paying users onto the API. By highest-paying users I mean the people paying for subscriptions. Hence the $100 credit a month. We helped train their products, and now we cost too much.
So where's the ethics in the supposedly most ethical frontier AI business of the bunch?
To Anthropic,
I'm a Max 5x subscriber, and I'm writing to complain about the changes to Claude since Chat and Cowork were merged into one app on 7 and 8 October. I use Claude for long-form investigative research and to build a complex software app, working in one project over long sessions, and since the merge it no longer does what your own help pages say it does. My chats compact far more often, with no warning, and the context window doesn't match your published figures. The chat history you say is “preserved so Claude can reference it” is being locked away from Claude as “flagged by safety classifier”, and nobody has told me. A safeguard flag fires when I copy thinking summaries that are already on my screen. The new API credits can't be spent in the app, which is where the extra usage is going. Your own postmortem in April admitted that changes like these make Claude forgetful, and they've been rolled out to paying subscribers anyway.
I've kept session logs and checked all of this against your documentation. The detail is below, and I want each question answered individually, in writing.
What context window am I actually getting?
Your help article “How large is the context window on paid Claude plans?” lists Opus 4.6 at 500K when chatting, 1M in Claude Code and 200K in Cowork. The same model runs at 1M in Code, so 200K is a setting you chose, not a limit of the model. Chat is Cowork now, so which number applies? GitHub issue #100502 was opened on 8 October by another Max 5x user. They measured about 92,000 tokens of fixed overhead before the first message and auto-compaction at about 142,600, which leaves roughly 50,000 tokens of working room per cycle. My own session log from 13 October shows almost the same: about 86,800 tokens of setup and compaction at 144,929. One Reddit user's breakdown of that overhead puts around 21,000 tokens on widgets for sports scores, recipes, shopping, travel, weather and quizzes. I'm doing research and building software. Why is my context paying for recipe cards?
Why does switching models wipe the chat?
On 13 October my chat had reached 676,481 tokens on Opus 5.5 without compacting. I switched to Opus 4.6 and within a minute it compacted down to 20,920. Three minutes later it compacted again, at 144,929, eleven seconds after it started reading my project files. That's two compactions in ten minutes and about seven minutes of waiting, and the first one processed the entire 676,481-token chat. I got no warning before either of them.
Where exactly is my history preserved?
The same article says “Your full chat history is preserved so Claude can reference it, even after earlier portions have been summarized.” On 13 October, and again on 15 October in a different chat, Claude tried to read back the compacted part of the conversation with its past-chats tool. Both times it got “[Content redacted, chat flagged by safety classifier]”. Since 28 September, 7 of the 19 chats I've used in my project have been locked this way. None of the 95 before that, going back to March, were. Nobody told me any of them had been flagged, what flagged them or when. The transcript the compaction summary points Claude to starts at the compaction itself and holds nothing from before it. So the history gets summarised, the summary points to a transcript that doesn't have it, and the original is locked away from the only thing that's meant to reference it.
Even the chat I wrote this letter in has been redacted. When I open a new chat, Claude can't refer back to this one, so all the work I've done in it is blocked from the next session. If the only way to cope with compaction is to keep starting new chats, and you keep redacting the old ones, how is anyone supposed to carry the work across into the new chat? Where does that leave the user?
What does this do to development work?
I'm building a complex app in the same project. A development project lives on its history: the decisions made, the approaches tried and rejected, the reasons behind the code, and the bugs already fixed. When a chat compacts, that history drops out of the live conversation, and when the chat is then locked, Claude can't get it back. The chat I was locked out of on 15 October held two days of design work and mock-ups for the app. So Claude suggests what we've already ruled out, undoes what was settled, and I pay usage to explain it all over again. Who builds anything serious on a tool that forgets the build?
What's the benefit to your loyal customers?
I want to know what this does for loyal, long-term, high-paying customers. How is it a benefit to get more compaction, and an AI that forgets the details of the work? The only workaround is to keep starting new chats, but nobody knows when compaction is coming, so you end up starting new chats constantly, working in fear of the next compaction, and trying to carry the work from one session into the next. There's no logic in it and there's no productivity in it. Who decided that forcing your heaviest, highest-paying users onto that would be productive? It's insane.
Why is a safeguard firing on what's already on my screen?
On 13 October my messages were stopped with “Opus 5.5's safeguards flagged this message” and the category “reasoning_extraction”. It fired on two more of my messages on 14 October. Your thinking docs describe that category as “Requests that try to pull the model's internal reasoning into the response text”. I wasn't doing that. I was copying the thinking summaries your app had already shown me into a record, and your docs say those summaries are written “by a different model from the one you target”. The same words went straight through when I typed them out by hand or sent them inside a text file. So the flag stops nobody who's determined, and all it gets in the way of is people keeping records of their own chats. Meanwhile your docs say “You're charged for the full thinking tokens generated by the original request, not the summary tokens.” I'm paying for reasoning I'm not allowed to see, and then getting flagged for keeping the summary of it.
Claude isn't told when these warnings come up, and it has no idea what they look like. So the only way it can know what's happening is if I show it. When the warning came up on 13 October, the app wouldn't let me send a screenshot of it into the same chat, and I couldn't copy its wording to paste in either. I had to type it out by hand to show Claude what was going on. What's the point of a warning the user can't show to the AI it's about?
You knew, didn't you?
Your postmortem of 23 April came eleven days after I published my own data on the March 26 change. It admitted that a change to clear Claude's older thinking, and a bug in that change, “made Claude seem forgetful and repetitive”. It called cutting Claude Code's default reasoning effort “the wrong tradeoff”, it listed Cowork among the products affected, and it said “We never intentionally degrade our models.” Six months later you've merged chat into Cowork, and the merged app compacts more often than ever and puts what it compacted out of reach. Pro and Max were moved first, while Enterprise admins were promised at least 30 days' notice. Issues #100502 and #100750 are both labelled as regressions, and as of 15 October neither had a reply from staff. When a regression you've already admitted to once comes back, is it still a mistake?
Why can't the credit be spent where the burn is?
Last week, for the first time ever, I maxed out my weekly usage before the week was over. This week I'm already at 50%, four days in. The same week as the merge, you started giving Max and Team plans monthly API credits, $100 on Max 5x. Your help page says they can't be used for “Extra usage in Claude, Claude Code, or Claude Cowork”, and that “Unused credits expire at the end of each billing cycle.” So the app burns through my limits faster, and the credit can only be spent by learning the API before it runs out. Running Opus 4.6 at 1M in Code needs usage credits switched on for Pro users. Is there any way out of this that doesn't cost more?
What's your end game?
I also find the timing interesting. On 8 October, while you were moving people onto the merged app, you published your 2026 Usage Policy update. It adds a prohibition on “sustained and needless abusive or cruel behavior toward our models” from 12 November, and the policy says that if you suspect a violation you “may warn you or throttle, limit, suspend, or terminate your access”. My account data shows the switch to the new setup on the same day that post went up. So you moved your heaviest users onto a setup that compacts constantly and loses the thread, which was always going to cause frustration and anger, and that same day you announced you can cut people off for how they talk to it. You say the rule doesn't apply to “common versions of user frustration”. Who decides where frustration ends? What is your end game here?
What do I want?
My account put back on the Chat setup I had before 7 October. You ran it for me right up until then, and you promised Enterprise customers at least 30 days' notice before moving them, so it exists and you can run it. I won't accept that it can't be done.
Credit for the usage I've lost since the merge, to compaction, to explaining things again and to locked chats.
The real context window for each model in the merged app, and a correction to whichever of your pages is wrong.
A visible context meter and a warning before compaction.
An explanation of why my chats are being locked as “flagged by safety classifier”, what flagged them, when, and why I wasn't told, and Claude's access to them restored.
A fix for the reasoning_extraction flag so it stops firing on people copying summaries that are already on their screen.
I want every question in this letter answered individually, by a person. A generic or copy-and-paste reply isn't an answer, and I'll treat it as no answer.
The Architect
Well written letter. Really hoping you will get an answer to questions about issues many of us are experiencing.